Artificial Intelligence has come a long way from being fictional to encompassing a wide range of categories from logistics, data mining, playing chess or diagnosing medical conditions, to name a few. Image generation has been one of the areas of research of several companies, and Nvidia’s AI seems to have achieved success in creating scarily realistic looking human faces.
Nvidia’s Tero Karras, Samuli Laine and Timo Aila have revealed the progress made by their AI in a new whitepaper. The results are actually shocking when considering the fact that the AI has generated faces of humans who don’t actually exist. Nvidia’s researches have built on top of the work of Ian Goodfellow from 2014 by introducing the principles of Style Transfer to a modified version of Generative Adversarial Network, also known as GAN. While GAN generates images after looking at a source, a Discriminator Network tries to separate the fake images from realistic ones.
To this, Nvidia’s researchers added style transfer principles to fuse the characteristics of one image into another. In essence, two neural networks combined with style transfer could generate some very realistic looking images of people who don’t even exist. Nvidia’s researchers tried to do the same with cat images too, but cats have proved to be difficult so far.
Is AI going too far?
Fake news has been a growing problem for the past few years. Deepfakes have become increasingly difficult to identify, and with Nvidia’s AI being able to generate images of non-existent people, this could pose bigger challenges in the future. There have been several demos of researchers being able to generate fake videos featuring popular politicians, and if this tech is abused, it can be used to create fake videos that look even more realistic.
That said, the world of cats may not have to worry about Nvidia’s AI. For now.